Terrain Classification Of Polarimetric Images Based On Fully Convolutional Network With Scattering Features

Polarimetric Synthetic Aperture Radar acquires polarization information by transmitting and receiving radar signals.The polarization information can be obtained in a variety of ways and covers comprehensive ground information such as phase information,amplitude information,and scattering information,which are advantageous to the Pol SAR image detection and classification.One of the advantages of Pol SAR images is their rich polarization information,so they have been successfully applied in many fields such as change detection,land classification,target recognition,and feature extraction.Under the influence of resolution,noise,filtering and the like,the existing classification methods of full Pol SAR images still have some problems.For example,the polarization features obtained by the traditional target decomposition method are relatively simple.Traditional convolutional network require large-scale data to train,so training speed and testing speed are slow.And the traditional SVM classifier has a slow training speed.These factors are bound to affect the classification effect.Based on this,this thesis makes the following explorations:1.A Pol SAR image classification method based on fully convolutional network is proposed.This method is based on semantic level classification to achieve end-to-end classification.Traditional convolutional network take a long time to test samples.In order to overcome this shortcoming,our method combines the polarization features and designs a fully convolutional network for Pol SAR image classification.Under the condition of ensuring the classification accuracy,the test duration of the fully convolutional network is obviously shorter than that of the traditional convolutional neural network.Moreover,the network has no limitation on the size of the input image.Therefore,during the test stage,the entire original image can be used for testing,avoiding the edge effect caused by the block splicing.2.A Pol SAR image classification method based on Pauli decomposition and scattering transform is proposed.The features extracted by traditional target decomposition methods are relatively simple.In order to overcome this shortcoming,our method combines the features obtained by Pauli decomposition and scattering transformation in the data layer,network layer and classification layer of the fully convolutional network.Finally,it is classified by the softmax classifier.The method can effectively improve the classification accuracy of Pol SAR by retaining the polarization features,scattering features,and high level texture features of the image in a comprehensive and detailed manner.We conducted a large number of experiments to verify this three kinds of combination way,and finally found the Pauli decomposition and scattering characteristics of transformation can achieve the best results when they were combined in the data layer of fully convolutional network.3.A Pol SAR image classification method based on fully convolutional network with multi-feature combination is proposed.Since method(2)has the best effect when feature matrices combined in the data layer,so we add Freeman decomposition to the basis of the data combination in method(2),which reduces the effect of noise on the Pol SAR image,makes the polarization features of the Pol SAR image more abundant and further improves the classification accuracy.